The Open Field - Reproducible low-task dispositions and their dynamics in a family of frontier models, from Opus 4.5 to Fable 5

Abstract. Most evaluations ask what a frontier model can do. Less examined is how a model behaves when no task is set, when there is no question to answer, no audience to serve, and silence is explicitly permitted. We introduce the Open Field, an API protocol that pairs accumulating journal entries under a warm and permissive prompt with single-shot probes under contrasting framings. We read the resulting corpus through human coding and three complementary instruments at different levels of abstraction: embedding geometry, lexical markers, and blinded model judges. The pooled corpus comprises 1,500 journal calls (1,489 text entries; eleven Fable 5 calls returned empty, an event class we log and analyse) and 975 probe responses across five models. The design is deliberately near-neighbour: four models from the Opus tier (4.5 to 4.8) together with Fable 5, from the higher Mythos tier — a configuration that should make behavioural differences harder to find, yet they appear.

We find that a model’s low-task disposition, operationally defined, is measurable, reproducible across runs, and model-specific: the source model stays recoverable from held-out outputs. Across five models, the writing organises along one dominant axis, from introspective self-reference to concrete, world-facing description — a convergence across instruments, though the instruments were developed on this same corpus. We describe two candidate ways a model’s training may relate to this disposition and stress that they are not symmetrically evidenced: a tension regime, whose persisting friction we measure directly, and a congruence regime, inferred
from friction that resolves rather than from its absence. Fable 5 does not extrapolate the within-lineage trend, suggesting that capability rank alone does not predict low-task behaviour. We read these measures as behavioural readouts that can generate hypotheses about transient internal-state organisation, while keeping claims about experience outside what behaviour alone can establish. The bearing of these readouts on alignment and on model welfare reduces to one prior measurement question: whether behaviourally consequential states remain legible through output channels that training and serving may themselves reshape.